A Study on the Impact of Hyperparameters on the Temporal Convolutional Network Model for Electric Load Forecasting

Electric load forecasting is essential for reliable power-system planning and operation. Temporal Convolutional Networks (TCNs) have gained attention for time-series prediction due to causal and dilated convolutions with residual connections, which help capture long-range temporal dependencies. However, TCN accuracy is highly sensitive to hyperparameter settings, and systematic evidence on which hyperparameters matter most for load forecasting remains limited. This study investigates major architectural and training hyperparameters in a TCN-based short-term load forecasting pipeline using 30-minute electricity-load data from New South Wales (NSW), Australia (2015-2021). The task is one-day-ahead forecasting with a 48-step horizon, and performance is evaluated primarily using Mean Absolute Percentage Error (MAPE) under a controlled experimental protocol with a fixed reference configuration. Results show that training hyperparameters-especially the learning rate-have the most substantial impact on accuracy and stability: a learning rate of achieves the lowest median MAPE of 2.375%, improving the reference configuration’s median MAPE of 3.212% by approximately 26% (relative). Architectural choices such as the dilation schedule and input window length have moderate effects, while encoder-decoder capacity parameters are comparatively less sensitive within the tested ranges. These findings provide practical guidance on prioritizing hyperparameters when configuring TCNs for short-term load forecasting, balancing forecasting accuracy and computational efficiency.